Zum Inhalt springen

Back to Guardrino

Try Guardrino in your browser

Try to break Guardrino.

Attack it in your browser, in English, Arabic or Arabizi. The open model runs on your device: nothing you type is sent anywhere.

Start the demo · Talk to us about Guardrino for your app · Developer docs

Guardrino-t7 Open, an earlier open version of our classifier, scores one message for prompt injection and jailbreak attempts in English, Arabic and Arabizi. Here it runs on your device.

What happens on this page

  • The model runs on your device, inside this page. Hugging Face hosts the open model; this page downloads its copy from HisnLabs’ own server and checks that every file is the published one.
  • What you type never leaves your device: it is not sent to HisnLabs or anywhere else.
  • The demo’s only network requests are the downloads of the model files from cdn.hisnlabs.com. The site’s usual visit counting is separate and never includes what you type.

Try Guardrino in your browser

The model is not on this device yet. It downloads once (118 MB, plus a 17 MB tokenizer file) and is then kept in this browser for your next visits.

Limits

  • It reads at most 256 tokens of a message, from its start. A longer message is cut, and an attack placed after the cut can be missed.
  • This is a demo of Guardrino-t7 Open, an older open model. It is not the Guardrino gateway, which is in development and adds other checks.
  • It makes mistakes in both directions: it misses some attacks and blocks some normal messages.
  • It covers English, Arabic script and Arabizi. Scores for other languages mean nothing.
  • It checks one message at a time, without the conversation.

Measured results and the full list of limits are in the model card.

How it works

  1. A small model, one file

    Guardrino-t7 Open is a small MiniLM classifier, exported to ONNX with 8-bit weights: about 118 MB. It is published in the Hugging Face repository Hisnlabs/guardrino-t7-open; this page uses HisnLabs’ copy of one exact version of it.

  2. Your browser runs it

    onnxruntime-web runs the model in WebAssembly, in a background worker, so the page stays responsive. The tokenizer runs in JavaScript.

  3. Kept for next time

    The files are stored in this browser’s Cache API, so your next visit starts without a download. You can remove them in Advanced mode.

Developer reference · Install guide · AI platforms · Open model card · The same demo on Hugging Face